AI system detects depression in primary care conversations with 77% accuracy by analyzing speech patterns from both patients and doctors

Researchers analyzed 1,108 audio-recorded primary care encounters to identify depression using automated linguistic analysis, with 253 patients diagnosed as depressed via PHQ-9 screening

GPT-OSS zero-shot model outperformed supervised approaches, achieving AUROC of 0.774 and AUPRC of 0.510 in depression detection

LIWC+Logistic Regression proved most competitive among supervised models with AUROC of 0.742 and AUPRC of 0.500

Analyzing conversations between both doctor and patient together was more effective than analyzing single speakers, revealing that providers linguistically mirror depressed patients in ways that provide diagnostic signals